Timeline for Is a DID model appropriate?
Current License: CC BY-SA 4.0
16 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 6, 2021 at 7:03 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Aug 18, 2020 at 18:00 | history | tweeted | twitter.com/StackStats/status/1295782550976237569 | ||
| S Aug 15, 2020 at 14:17 | history | suggested | Thomas Bilach | CC BY-SA 4.0 |
Edited the text. Added appropriate tags.
|
| Aug 15, 2020 at 1:59 | review | Suggested edits | |||
| S Aug 15, 2020 at 14:17 | |||||
| Aug 14, 2020 at 20:20 | comment | added | Thomas Bilach | Nothing is impossible. But you will have to make some ad hoc decisions if you're going with a difference-in-differences estimation strategy. Note, one-third of all U.S. states do not have sufficient pre-policy data. You could drop these from your sample, but you're getting rid of a lot of observations. | |
| Aug 14, 2020 at 19:30 | answer | added | Thomas Bilach | timeline score: 2 | |
| Aug 14, 2020 at 19:02 | comment | added | Bryan | I was wondering about the "enough data" thing. How else could I look at it? Is it just impossible to analyze? | |
| Aug 14, 2020 at 18:23 | comment | added | dimitriy | Synthetic Cohort approach might be doable with all three groups (depending on the timing details), but you will have to try it to know for sure. I also don't know of an R package that can handle multiple treatments with variable start dates. | |
| Aug 14, 2020 at 18:23 | comment | added | dimitriy | You are interested in the distribution of counterfactual outcomes. You don't observe the treated outcomes for untreated units and vice versa, so there is sampling happening, so your SEs reflect that uncertainty. It will be hard to make use of the always treated group with DID. DID can be used with repeated cross sections, but that assumption seems hard to swallow with state-level data. Without those, I really wonder if you have enough data for a DID analysis. | |
| Aug 14, 2020 at 16:34 | comment | added | Bryan | So, then, it is not possible under any circumstances to compare entire states, since data for all 50 states is the ENTIRE population. "Sampling" and "randomization" are impossible. Therefore, it is impossible to compare the states to each other. Interesting premise. As for your rainfall/graduation comparison, you showed an association, but no association every proves anything in any mechanistic manner--ever. It only shows an association. It may speak to a latent variable to which rainfall is a contributor, though. Ever look into wealth of a state vs. rainfall? | |
| Aug 14, 2020 at 16:31 | comment | added | BruceET | I am having trouble viewing this as a situation for a valid significance test. Where is the sampling ? Where is the randomization? How about a table or colored map describing your data. // I once "showed" that among US states higher annual rainfall "predicts" lower high school graduation rates. | |
| Aug 14, 2020 at 16:02 | comment | added | Bryan | So I test a model like "outcome ~ year + policy + year x policy" where the "year x policy" term is the DiD term and "policy" is a binomial variable indicating each year whether or not the specific state has the specific policy? | |
| Aug 14, 2020 at 15:49 | comment | added | Bryan | I should say "eighteen states enacted in different years within the period". | |
| Aug 14, 2020 at 15:40 | comment | added | Bryan | Yes, exactly. I have 1984-2018 annual outcomes data (income, employment, inflation, all by state). Nineteen states enacted before 1984. Eight states enacted within the period. The remainder don't have it. | |
| Aug 14, 2020 at 15:19 | comment | added | Thomas Bilach | This could be modeled using difference-in-differences. When you say “group” you are, in fact, referring to a subset or collection of U.S. states, correct? Also, how many times periods do you have? | |
| Aug 14, 2020 at 14:58 | history | asked | Bryan | CC BY-SA 4.0 |